PUB - Publikationen an der Universität Bielefeld

We propose Contour-Net as a bio-inspired model for rhythmic movement control of a pair of insectoid feelers, able to successively sample the contour of arbitrarily shaped objects. Initial object contact initiates a smooth transition from a large-amplitude, low-frequency searching behaviour to a local, small-amplitude and high-frequency sampling behaviour. Both behavioural states are defined by the parameters of a Hopf Oscillator. Subsequent contact signals trigger a 180 phaseforwarding of the oscillator, resulting in repeated sampling of the object. The local sampling behaviour effectively serves as a contour-tracing method with high robustness, even for complicated shapes. Collected contour data points can be directly fed into an artificial neural network to classify the shape of an object. Given a sufficiently large training data-set, tactile shape recognition can be achieved in a position, orientation and size -invariant manner. Only minimal pre-processing (normalisation) of contour data points is required.